SoundCloud Widget Placeholderhttps://soundcloud.com/uclsound/ep-5-robotics-and-simulation-for-or-agai... Episode 5: Robotics and Simulation: For or Against HumanityCassidy 00:03
Hello, and welcome to IM@UCL: The Podcast, a podcast about the research at UCL that will revolutionise the future of driving. My name is Cassidy Martin and I am your host on this journey of self-driving discovery. Nowadays, you can find robotics and simulations everywhere. Rover vacuums and mops can be found in households, virtual reality headsets are used to play games, and touch sensitive prosthetic devices are giving people new limbs to use. The regular utilisation of robotics and simulations will only continue to increase over time. And although these devices open up a world of possibilities, it is important that their creation and integration is treated with caution. We will get into why this is later on. For this month's episode, I spoke with two academics who are focused on utilising robotics and simulations to create a better world for us all. Our first guest talks about how data on human behaviour can optimise the way we commute. And the next, discusses using simulations to teach semi-autonomous vehicles how to drive. Let's get started. Tim 01:21
My name is Tim Hillel, I'm a lecturer in the Infrastructure Systems Institute and Centre for Transportation Studies at UCL. Cassidy 01:29
Tim leads the Behaviour and Infrastructure Group at UCL, which aims to build a deeper understanding of cities and infrastructure systems using human behaviour as the key part of analysis. Their current research focus is developing powerful simulations that model how people move through urban areas that can then be used to help optimise the design and operation of those areas. Tim 01:53
When we do simulations of urban systems, we need to basically model the choices and behaviour of a population. And so, we have different sources of data describing the population. But either we can't use that data directly, so in the case of something like the census data, it's too tightly, we can't just let everyone use only the census data; or we might only have data about a small portion of the population we're interested in. So, we might do a survey and get some information about the mobility of a subgroup of people in London. And the idea of creating synthetic populations is that may be the small sample data that we have, but also preserve potentially privacy of other datasets. So, one example that we can use synthetic data for is direct use and simulation, so creating these synthetic populations. But there's other reasons you might want to do it as well. Improving the training of machine learning algorithms, for instance, is another avenue that we're looking at. Cassidy 02:44
To build simulations of urban areas, Tim makes use of synthetic data, which he generates using artificial intelligence (AI) models. Tim 02:56
At a basic level, what we want to simulate is people's behaviour. And we want to simulate how they interact with the city. And so that requires basically two things, we first need a model that represents their behaviour. But then we also need a kind of representation of the people themselves. And so it's that synthetic population ¬– or you might have heard of an agent based model, the agents in that model ¬– that's what we model with, this kind of synthetic data generation. And so we try to create those from existing data sources. And the aim is to make them as realistic as possible. So that they represent quite well, the city that we're looking at or the urban area. Cassidy 03:35
And when you're doing this, does that include all kinds of different types of transport? Or are you looking at specific forms? Tim 03:44
Yes, so the kind of activity-based approach we're looking at, so the way this works, is we simulate the different activities people do in a day. If, for instance, they go to work, they might from work then go and get lunch outside work, they might do a social activity after work – and we simulate their choices to do these activities and to schedule these activities. And then as part of that, we model their desire to travel to those activities. And so, in your working day, you might take public transport to get to work. But then that means that for tours that you do from work, so like when you go to your lunch, you might choose a different mode. And so, you might then walk to lunch, and so on. So, the idea is to build up this picture of how people travel in a day based on the activities they partake in. And it's multimodal, yeah. Cassidy 04:37
And what can you, what are you hoping to do with all of this information? So, having these synthetic models and knowing what these people are doing? Tim 04:47
Yeah, that's a good question. You always want to be doing something with your model. The main idea is to investigate things that haven't happened or might happen in future. So, the idea is, if your simulation can give you reliable results about what people would do under different scenarios. It would allow us to investigate scenarios that we've not seen yet. And so, these kinds of counterfactual scenarios as we call them, are these what if questions. So, what if I build a new station here? What if I allow people to work from home? What if I introduce autonomous vehicles? And how does that affect people's decisions? In terms of how they travel, but also how they schedule their days. So, the idea is really kind of long term forecasting. But there's also possibilities to use these kinds of approaches for like the here and now, what's going on in a city. Cassidy 05:39
And these simulations that Tim is creating can help in areas beyond research and semi-autonomous vehicles. Tim 05:48
Everyone's trying to optimise something. And so, I hoped to be working more with people that were looking to optimise their environmental footprint, and encouraging people to take more environmentally friendly modes of travel. But you know, the possibilities of what you can do once you have an accurate city simulator are quite, I guess, exciting. And it's not just limited to mobility. So, we could, we could also look at things like we can use this utility to represent people's satisfaction with their days, and so on. And I think with the big disruption, we're seeing with technology and COVID, there's an opportunity for us to really rethink the way that cities work. And I would hope that by having these digital representations of how cities work, we can test out lots of ones and find out which ones actually work best. So designing our cities around more than just what's the obvious next step for what we construct. But actually, you know, is there a better way of us operating as a city? Is it more efficient and more friendly and more, a nicer place to be? Cassidy 06:49
And this idea of using synthetic data to create realistic simulations of urban areas fits in nicely with the research facilities available at PEARL’s IM@UCL. Tim 07:02
So yeah, so I think what's really exciting at UCL in particular, and what IM plays a part in is this idea that we can control the environment and observe how people behave in that environment. So, with this new research facility, PEARL, it opens up all sorts of opportunities for us to create models with the bits that are usually kind of left aside in these representations. So, we can investigate things like what actually are the impacts of comfort, of warmth, of accessibility? And I think where this opens up new opportunities is then how do we then turn those kinds of experiments in the sandbox into representations that we can then use in simulation for investigating these scenarios? And obviously, with this new simulator, and the tools that IM has, I think the real exciting thing is to see what potential impacts might self-driving cars, or you know, higher autonomy of vehicles have on the choices that we make. So, for instance, you know, if I don't have to pay attention when I'm driving, can I productively work? Do I have a mode of transport that's like the train in that I don't need to focus, but like the car, in that I don't need to interact with other people and I can get to my destination very quickly. And so yeah, if we can start to do that, then we can potentially be some of the, you know, researchers that can answer questions that no one can really answer yet. Which is like, well, how many people are likely to use these cars? Or different modes of transport, or mobility service, or whatever it is? And how will that impact the way that people interact with cities? So, I think that's what's really exciting, is this idea that we can really. the facilities that UCL allow us to investigate things that most other people just have to kind of put down as a constant in the model. Okay, yeah. We think that this train is more comfortable, but we can actually say, well, let's create a train and see. Cassidy 08:54
That's cool. So, I guess, I can think of why this is, but why do you think this type of research is important? Tim 09:02
So the sales pitch answer is we are all going through massive changes and infrastructure is going through a huge change. And actually, in reality, it's true, right? We, we have the technical, technological innovations that we're seeing are just causing real step changes in the way that people behave in cities. And some of them are much, not necessarily, the kind of technologies we think of first. But even just ride hailing, right? So, ride hailing apps have had a huge impact on the way that people particularly use night buses. And this causes like, overnight issues, ride sharing gets introduced in the city, and suddenly the night buses don't make sense anymore. And then, I mean, we've had this massive impact on life from COVID, which is ongoing, but it's, it's also opened up all these, I guess, possibilities of homeworking. And so, we're gonna see much, much higher rates of people working from home or working in unconventional locations. There's whole industries, not just companies, but whole industries which are just deciding to go permanently to remote working. And yeah, I mean, if we're planning and structure around this, we really need to understand what impacts that's going to have. I think that's why the counterfactuals are so exciting because we can envision what might happen and see if we can test that. Cassidy 10:15
Yeah. I guess with your area, you want to be familiar with all forms of transportation and all future possible forms of transportation as well so that's why automated cars are good to know. Tim 10:29
Exactly, there's all these, I mean, one of the things that's very interesting about the UK compared to other countries in Europe is, I mean, once an idea is a concept, there's a lot of push to make it kind of a thing. Before we've got standardised digital twins, we talk about the national digital twin and we have this ambition to build a national digital twin and… Cassidy 10:46
I don’t know what a digital twin is. I'm sorry, what is it? Tim 10:49
So digital twin is when you build a digital representation of anything that somehow connected to the physical anything. And so, when we do it in cities, or infrastructure, we build a digital twin representation, so a digital model of a city. And then we try and link it to the city with real time data, and so on. So that can use our digital twin to investigate. Sometimes could be something like, where are people at this moment? But we'd also start to, if we can simulate within that twin, we can look at these what happens if I change the network slightly? Or what happens if the conditions change? So yeah, it's a big buzzword at the moment, digital twins and digital twin cities. Cassidy 11:30
Cool. And so I just have one last question for you. So, what are you most hoping to achieve with this research? Tim 11:37
So yeah, what I think would be really exciting, and I guess the kind of long-term goal, is coming up with a way for us to rethink the way in which cities work to design them to be best for the people that actually use them. Not just kind of top down philosophy, but just in terms of actually how efficient a city can be both in terms of, you know, getting people around with low footprint and using an infrastructurally low footprint, but also in terms of how they are as places to live and exist in and interact with. And so if we can do that in a in a data driven way, I think that's what could be the most exciting. Cassidy 12:16
Now after learning about the creation of simulated cities, we're going to switch gears and find out about a researcher whose work in semi-autonomous vehicles could benefit from simulated cities. Just a note before we start: for the comfort of our next guest, the answers to my interview questions in the following segment were re-recorded by him and have not been edited. Dimitrios 12:43
My name is Dimitrios Kanoulas, Associate Professor in Robotics at UCL, department of Computer Science. My work is on perception and learning for complex robots, such as humanoids and quadrupeds. Cassidy 12:57
Your brain, like mine, might immediately think of robots you've seen in sci-fi films and TV shows. But the robots Dimitrios works with serve a higher purpose than media entertainment. Dimitrios 13:11
A few years back in central Italy, after a severe and disastrous earthquake, we had a humanoid robot entering a destroyed building to assess cracks and extract some objects. We asked the mechanical and civil engineers to be completely outside the destroyed building and supervise the inspection from distance, so that they do not risk their lives, while the robot was operated by our researchers under their supervision. Inspection of destroyed buildings is just an example of tasks that are very dangerous for humans to complete. In the case of post-earthquake inspections, traditionally engineers are assessing each affected building. Somebody must go inside and inspect even if a building is red listed. What if the building collapses? The question is: could we use robots to help humans and avoid such dangerous scenarios? Another example. What happens when there is an abandoned suspicious bag in an airport? In this case, specialized humans come and try to figure out if there are explosives so that they can dispose them. There are thousands of such cases per year in the UK, and millions worldwide. Could robots help in this case? Instead of risking human life, we thought of allowing a specialist to control a relatively cheap robot from distance, approach the suspicious object and tele-operate the robot with precision to dispose the explosives, if any. By using a virtual reality headset, the specialists can feel as being at the scene and operating themselves, but in reality, without risking much. One risks only the robot in case something goes wrong. These are two cases of essential tasks, in which robots could protect human life. Cassidy 15:29
Not only can these complex robots complete tasks that are dangerous, they can also complete tasks that are monotonous. Dimitrios 15:36
Imagine a task that is repetitive and tedious, but also important and potentially dangerous. For example, what if you had to go inside a nuclear power plant and check the levels of radiation at different areas of the plant? Having a human completing this task every day for years is tedious, dangerous, and after all kills human creativity. What if instead we had a robot that can automatically complete this type of tasks without risking any human life? Quadruped robots can do that nowadays, of course after some human supervision. In the meantime, jobs are not lost. Humans are instead remaining safe and are there to supervise the tasks from distance, as well as fixing and coding the robots. Our intension is to help with dangerous and tedious tasks – not robots to conquer over human workers.
Cassidy 16:35
I think it's important to talk about, or people to hear about, those positive sides of it because there are so many like things in the media that make you kind of nervous, or movies or maybe even like, I don't know, I'm sure you get this a lot, the Black Mirror episode? If you're wondering what I am referring to here, it's an episode of Black Mirror called Metalhead where someone is chased by a terrifying robotic guard dog. Dimitrios 17:02
I do understand this. People should realise that we, researchers, think about this. Technological advancements are not by default good or bad. It depends on their purpose and use. In robotics, we already talk about trustworthiness and ethics, areas that explore those questions. Imagine an extreme case. Can you place a gun on top of an autonomous legged robot? The answer is no, and our community tries a lot to block the use of robots for such use-cases. Even beyond the ethical part, there is always the question of trustworthiness. This is an extreme example, but very recently some of the most successful companies that sell legged robots signed a letter against weaponizing their robots. It is true that we should carefully think of the use of the robots and the methods we develop. Nothing should be taken for granted. Everything should be discussed in our community, including ethics and trustworthiness. Questions such as: What is the purpose and result of using this robot for that application? It is up to us to make Robotics and AI for good.
Cassidy 18:18
Yeah. And who would be like the people kind of policing this? The United Nations or something? Or like, I don’t know. Dimitrios 18:24
That's a good question. Robot ethics is an active area of study and research. UN or a similar organization could play this role globally, but efforts need to be made at all levels, from whole countries to small labs, in order regulations to be generated. We can protect our world and we can make sure that robots will not be used against humanity or nature.
Cassidy 18:51
Just as making sure humanoids, robot dogs, and the like are designed to be safe, semi-autonomous vehicle robots also need to be designed to be safe. And these robots can learn how to be safe from experience, such as using the type of driving simulator found at IM@UCL’s facility. Dimitrios 19:11
Let's say that you would like to develop an autonomous car. And what you want to do is make it drive around the city on its own. How could you do that? One could really have a driver driving around the city and try to replicate the driver’s actions. What if you would like to learn about tricky and risky cases, where a car stops suddenly in front of you, or a human crossing unexpectantly the road? Real life driving cannot risk being involved in such cases, in order develop methods for an autonomous car driving. A simulator plays this exact role. Instead of really driving around the city, and risking human life or causing accidents, one could drive in a simulated city and generate as many scenarios as possible either standard or risky ones. Machine learning methods could take all the data as input to generate autonomous driving models, which will then be applied on the physical robot. Of course, this will require some fine tuning and real-world re-training, as simulators might not be able to replicate in full the physical world and, still, human supervision might be needed with some type of “emergency stop”.
Cassidy 20:33
That makes sense, because you can't always predict everything from a simulator, there's always going to be something. Even if it isn't a perfect system, it's better to have some kind of simulator to be able to test with than have nothing at all. Yeah, and having that emergency stop. I guess, you know, and that's probably why we won't go completely autonomous I imagined for a while with cars. Well, do you think they'll always have an emergency stop? Or do you think they'll get to a point where they won't need it anymore? Dimitrios 21:04
In my opinion, really depends on the robustness and safety of the system. Right? This comes with theoretical proofs, experiments, and real-world trials. Let’s take a different example. Airplanes. Autopilot is a reality there. Although still the take-offs and landings are manual. Why? Because these are the most tricky and dangerous parts of flying an airplane and full safety and robustness are not achieved yet. In autonomous driving, it is the same. Roads and traffic, for instance, are very complex systems. They are hard to be modelled and thus hard to generate methods to deal successfully in all possible cases. Sometimes, there is not even a single driving reaction to a particular scenario. Autonomous driving takes time, it really takes time to be developed. And therefore, in autonomous vehicles, it will take time to make them work without a supervisor there. This is very similar to our research on legged robots. When no humans or animals are around, it is safer to allow a fully autonomous robot inspect a building. But what if humans are there? Supervisors are needed to stop the robot in case a human is in danger. When we do research, we always have a way to stop the robot safely and not harm the operators. This is essential. Cassidy 22:37
So it might be a while before we go completely autonomous in driving (and flying). In the meantime, researchers like Dimitrios are making sure that every step towards autonomy is also a step towards safer driving. Thank you for listening to IM@UCL: The podcast. If you would like to learn more about the research and IM@UCL, you can check out their website at www.ucl-intelligent-mobility.com and/or subscribe wherever you are listening to this podcast so you can be notified when new episodes come out. This episode was produced and hosted by myself, Cassidy Martin, with music from Blue Dot Sessions. It was brought to you by IM@UCL, which is part of UCL PEARL in Dagenham, and supported by UCL Minds, bringing together UCL knowledge, insight and expertise through events, digital content, and activities that are open to everyone. A special thank you to Tim and Demetrius this month for sharing their time, knowledge, and insight. I hope you enjoyed listening to this podcast and feel like you learn something new, like I have with everyone I've interviewed in this series. Take care. And I'll see you again next month. Same time, same place. Cheers! |
Close
